Page 105 - Introduction to AI Robotics

88
3 Biological Foundations of the Reactive Paradigm
Pole vaulters also make minute adjustments in where they plant their pole
as they approach the hurdle. This is quite challenging given that the vaulter
is running at top speed. It appears that pole vaulters use optic flow rather
than reason (slowly) about where the best place is for the pole. (Pole vaulting
isn’t the only instance where humans use optic flow, just one that has been
well-documented.)
In most applications, a fast computer program can extract an affordance.
However, this is not the case (so far) with optic flow. Neural mechanisms in
the retina have evolved to make the computation very rapid. It turns out that
computer vision researchers have been struggling for years to duplicate the
generation of an optical flow field from a camera image. Only recently have
we seen any algorithms which ran in real-time on regular computers. 48 The
point is that affordances and specialized detectors can be quite challenging
to duplicate in computers.
Affordances are not limited to vision. A common affordance is knowing
when a container is almost filled to the top. Think about filling a jug with
water or the fuel tank of a car. Without being able to see the cavity, a person
knows when the tank is almost filled by the change in sound. That change
in sound is directly perceivable; the person doesn’t need to know anything
about the size or shape of the volume being filled or even what the liquid is.
One particularly fascinating application of affordances to robotics, which
also serves to illustrate what an affordance is, is the research of Louise Stark
and Kevin Bowyer. 135 A seemingly unsurmountable problem in computer
vision has been to have a computer recognize an object from a picture. Liter-
ally, the computer should say, “that’s a chair” if the picture is of a chair.
STRUCTURAL MODELS The traditional way of approaching the problem has been to use structural
models. A structural model attempts to describe an object in terms of physical
components: “A chair has four legs, a seat, and a back.” But not all chairs fit
the same structural model. A typing chair has only one leg, with supports
at the bottom. Hanging baskets don’t have legs at all. A bench seat doesn’t
have a back. So clearly the structural approach has problems: instead of one
structural representation, the computer has to have access to many different
models. Structural models also lack flexibility. If the robot is presented with a
new kind of chair (say someone has designed a chair to look like your toilet
or an upside down trash can), the robot would not be able to recognize it
without someone explicitly constructing another structural model.
Stark and Bowyer explored an alternative to the structual approach called
GRUFF. GRUFF identifies chairs by function rather than form. Under Gibso-
nian perception, a chair should be a chair because it affords sitting, or serves